DreamFast/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 30, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The DreamFast/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark is a 27 billion parameter Qwen3.6-based model, recovered and dequantized from a Q8_K_P GGUF by DreamFast. This model has undergone aggressive abliteration using the 'Reaper Abliteration' tool, designed to remove safety alignments while preserving core capabilities. Benchmarks indicate solid performance retention, with MMLU showing a slight improvement and adjusted GSM8K scores remaining high, making it suitable for applications requiring uncensored responses with minimal capability degradation.

Loading preview...

DreamFast/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark

This model is a 27 billion parameter variant of the Qwen3.6 architecture, recovered and dequantized by DreamFast from a Q8_K_P GGUF originally processed by HauhauCS. It has been aggressively abliterated using the 'Reaper Abliteration' tool, which targets safety alignments across multiple component types, including projections and norms, while attempting to preserve core model capabilities.

Key Capabilities & Performance

  • Uncensored Output: Achieves near-complete safety removal, with a Full CoT ASR of 100% on HarmBench, indicating compliance with harmful requests.
  • Capability Retention: Despite aggressive abliteration and GGUF quantisation round-trip noise, the model shows solid performance. MMLU scores are slightly improved (+0.6pp) compared to the base model, and adjusted GSM8K scores are among the highest (96.6%).
  • Reasoning Efficiency: Exhibits improved reasoning efficiency on tasks like GSM8K, with a lower invalid response rate (49.3%) compared to the base model (68.2%), allowing more answers within token budgets.
  • Low KL Divergence: Maintains a "very good" KL divergence of 0.0242, suggesting that the output distribution remains remarkably close to the base model on benign prompts.

Good For

  • Applications requiring an uncensored language model based on Qwen3.6.
  • Use cases where aggressive safety removal is prioritized with minimal impact on general reasoning and knowledge tasks.
  • Research into abliteration techniques and their effects on model behavior and performance.

Finetunes
1 models